Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A recognition system comprising: a sensing unit configured to perform sensing to output a sensor value; a task-specific unit including a first recognition processing part that performs a first recognition task based on the sensor value and a second recognition processing part that performs a second recognition task based on the sensor value; and a generic-feature extraction unit including a generic neural network disposed between the sensing unit and the task-specific unit, the generic neural network being configured to receive the sensor value as an input to extract a generic feature to be input in common into the first recognition processing part and the second recognition processing part, wherein the generic-feature extraction unit is connected to both the first recognition processing part and the second recognition processing part, and the generic feature is commonly used in both the first recognition processing part and the second recognition processing part.
2. The recognition system according to claim 1 , wherein the generic-feature extraction unit is provided on a semiconductor chip different from a semiconductor chip on which the task-specific unit is provided.
The invention relates to a recognition system designed to process data by extracting generic features and performing task-specific recognition. The system addresses the challenge of efficiently handling diverse recognition tasks while maintaining flexibility and performance. The system includes a generic-feature extraction unit that processes input data to extract features common across multiple recognition tasks, such as image, speech, or sensor data. These extracted features are then passed to a task-specific unit, which is optimized for a particular recognition task, such as object detection, speech recognition, or gesture recognition. The task-specific unit processes the generic features to generate a final recognition output. A key aspect of the invention is the physical separation of the generic-feature extraction unit and the task-specific unit onto different semiconductor chips. This separation allows for modular design, enabling the system to adapt to different tasks by swapping or updating the task-specific unit without modifying the generic-feature extraction unit. The separation also facilitates parallel processing, where the generic-feature extraction unit can process data for multiple tasks simultaneously, improving overall system efficiency. The system is particularly useful in applications requiring real-time processing, such as autonomous vehicles, robotics, and smart devices, where both performance and flexibility are critical.
3. The recognition system according to claim 2 , wherein the generic neural network in the generic-feature extraction unit includes hardware on the semiconductor chip.
The invention relates to a recognition system designed to process and analyze data, particularly for tasks such as image or pattern recognition. The system addresses the challenge of efficiently extracting and processing features from input data using a neural network architecture. A key component is a generic neural network integrated into a generic-feature extraction unit, which is responsible for transforming raw input data into a standardized feature representation. This feature extraction process is crucial for subsequent recognition tasks, as it enables the system to identify and categorize patterns within the data. The generic neural network is implemented using hardware components on a semiconductor chip, ensuring high-speed processing and energy efficiency. This hardware integration allows the system to perform complex computations required for feature extraction without relying solely on software-based solutions, which can be slower and less power-efficient. The use of a generic neural network means the system can be adapted to various recognition tasks by adjusting the network's parameters or architecture, rather than requiring a completely new design for each application. The system is designed to be modular, allowing the generic-feature extraction unit to interface with other processing units or algorithms. This flexibility ensures that the recognition system can be integrated into broader applications, such as autonomous systems, medical imaging, or industrial automation, where efficient and accurate data recognition is essential. The hardware-based implementation of the neural network further enhances the system's reliability and performance, making it suitable for real-time applications.
4. The recognition system according to claim 1 , wherein the first recognition processing part includes a neural network for the first recognition task that receives, as an input, the generic feature output from the generic-feature extraction unit to output a result of the first recognition task.
The invention relates to a recognition system designed to perform multiple recognition tasks, such as object detection, classification, or segmentation, using a shared feature extraction process followed by task-specific recognition modules. The system addresses the inefficiency of traditional approaches that require separate feature extraction for each recognition task, leading to redundant computations and increased processing time. The system includes a generic-feature extraction unit that processes input data, such as images or video frames, to generate a set of generic features. These features are then provided to multiple recognition processing parts, each specialized for a different recognition task. One of these processing parts includes a neural network specifically trained for a first recognition task, such as object detection or classification. The neural network takes the generic features as input and produces a result for the first recognition task, such as identifying objects within the input data or assigning class labels. By sharing the generic-feature extraction unit across multiple recognition tasks, the system reduces computational overhead and improves efficiency compared to systems that process each task independently. The neural network in the first recognition processing part is optimized for its specific task, ensuring accurate and reliable results while leveraging the shared features. This modular design allows for scalability, enabling additional recognition tasks to be integrated without redesigning the entire system.
5. The recognition system according to claim 1 , wherein the sensing unit includes a sensor that acquires the sensor value and a preprocessing part that performs preprocessing to the sensor value.
A recognition system is designed to process sensor data for identifying or classifying objects, events, or patterns. The system addresses challenges in accurately interpreting raw sensor data by incorporating preprocessing steps to enhance data quality and reliability before analysis. The sensing unit in this system includes a sensor that captures sensor values, such as measurements from optical, acoustic, or environmental sensors, and a preprocessing part that processes these values. The preprocessing part may apply noise reduction, normalization, filtering, or other techniques to improve the signal-to-noise ratio and extract relevant features. This preprocessing step ensures that the subsequent recognition algorithms receive clean, standardized data, leading to more accurate and consistent recognition results. The system may be used in applications like object detection, speech recognition, or environmental monitoring, where raw sensor data often contains noise or variability that must be mitigated for reliable performance. By integrating preprocessing directly into the sensing unit, the system streamlines data handling and improves overall recognition accuracy.
6. The recognition system according to claim 1 , wherein the generic-feature extraction unit includes a discrete device that resolves the input to each layer of the generic neural network, into integer bases.
Image and data recognition systems. The problem addressed is the efficient and accurate extraction of generic features from input data for processing by neural networks. The described system utilizes a generic-feature extraction unit. This unit contains a discrete device. This discrete device is configured to resolve the input provided to each layer of the generic neural network into integer bases.
7. The recognition system according to claim 1 , wherein the generic neural network has an integer weight.
Image recognition system. Addresses the need for efficient and potentially hardware-accelerated image recognition by incorporating integer-weighted neural networks. This system comprises a generic neural network designed for image recognition tasks. A key feature is that the weights within this neural network are represented as integers, not floating-point numbers. This integer representation can reduce computational complexity and memory footprint, making it suitable for embedded systems or scenarios where floating-point arithmetic is less desirable or computationally expensive. The system leverages this integer-weighted neural network for performing image recognition.
8. The recognition system according to claim 1 , wherein the generic-feature extraction unit includes a discrete device that resolves the input to each layer of the generic neural network, into integer bases, and the generic neural network retains a weight discretized having binary numbers or ternary numbers, the generic neural network being configured to perform the entirety or part of internal computing with a logic operation, the generic neural network being configured to transform a result of the logic operation with a non-linear activating function, the generic neural network being configured to give a result of the transformation to a next layer.
The invention relates to a recognition system that uses a generic neural network for processing input data. The system addresses the challenge of efficiently implementing neural networks in hardware by reducing computational complexity and memory requirements. The generic-feature extraction unit within the system includes a discrete device that decomposes the input data into integer bases before feeding it into the neural network. The neural network itself uses weights that are discretized into binary or ternary numbers, allowing computations to be performed using logic operations rather than traditional arithmetic. This approach simplifies the hardware implementation, as logic operations are more efficient and require less power. The neural network also applies a non-linear activation function to the results of these logic operations, enabling it to model complex patterns. The transformed output is then passed to the next layer for further processing. This design reduces the computational overhead while maintaining the network's ability to perform accurate recognition tasks. The system is particularly useful in applications where low-power and high-efficiency processing is critical, such as edge devices or embedded systems.
9. The recognition system according to claim 1 , wherein the generic-feature extraction unit includes a communication module or is connected to the communication module, the generic-feature extraction unit being configured to update the weight of the generic neural network, based on information received by the communication module.
A recognition system is designed to process and analyze input data, such as images or signals, to extract and recognize features. The system includes a generic-feature extraction unit that uses a generic neural network to identify features from the input data. This unit is equipped with or connected to a communication module, allowing it to receive external information. The system is configured to update the weights of the generic neural network based on the information received through the communication module. This enables the neural network to adapt and improve its feature extraction capabilities over time, enhancing the accuracy and reliability of the recognition process. The communication module may facilitate updates from remote servers, user feedback, or other data sources, ensuring the system remains current and effective in dynamic environments. This adaptive learning mechanism is particularly useful in applications where the input data varies or evolves, such as in surveillance, medical imaging, or autonomous systems.
10. A generic-feature extraction unit comprising: a generic neural network disposed between a sensing unit and a task-specific unit, the sensing unit being configured to perform sensing to output a sensor value, the task-specific unit including a first recognition processing part that performs a first recognition task based on the sensor value and a second recognition processing part that performs a second recognition task based on the sensor value, the generic neural network being configured to receive the sensor value as an input to extract a generic feature to be used in common between the first recognition processing part and the second recognition processing part, wherein the generic-feature extraction unit is connected to both the first recognition processing part and the second recognition processing part, and the generic feature is commonly used in both the first recognition processing part and the second recognition processing part.
A generic-feature extraction system is designed to improve efficiency in multi-task recognition systems by extracting shared features from sensor data. The system includes a generic neural network positioned between a sensing unit and multiple task-specific recognition units. The sensing unit captures sensor data and outputs sensor values. The task-specific units perform distinct recognition tasks, such as object detection or classification, using the same sensor input. The generic neural network processes the sensor values to extract a common set of features, which are then shared between the different recognition tasks. This shared feature extraction reduces redundancy, as the same features are reused across multiple recognition processes, improving computational efficiency and performance. The system ensures that the extracted features are applicable to both recognition tasks, allowing for flexible and scalable multi-task learning. This approach is particularly useful in applications where multiple recognition tasks must be performed simultaneously, such as in autonomous systems or advanced robotics.
11. A recognition system configuration method of configuring the recognition system according to claim 1 , the recognition system configuration method comprising: causing the generic neural network to learn with, as learning data sets, input data and output data of a learned recognition device that performs the first recognition task and input data and output data of a learned recognition device that performs the second recognition task.
This invention relates to a method for configuring a recognition system that leverages a generic neural network to improve recognition tasks. The system addresses the challenge of training a neural network to perform multiple recognition tasks efficiently by utilizing pre-trained recognition devices. The method involves configuring a recognition system by training a generic neural network using input and output data from two different learned recognition devices. Each recognition device is specialized in performing a distinct recognition task, such as image classification, speech recognition, or object detection. The generic neural network is trained on the combined input and output data from these devices, enabling it to generalize across multiple recognition tasks. This approach allows the system to adapt to new tasks without extensive retraining, improving efficiency and performance. The method ensures that the generic neural network can effectively handle diverse recognition tasks by learning from the specialized outputs of pre-trained devices, reducing the need for task-specific training data. The system is particularly useful in applications requiring multi-task recognition, such as autonomous vehicles, robotics, and smart surveillance systems.
12. The recognition system configuration method according to claim 11 , wherein an ensemble recognition device that unifies recognition results of a plurality of recognition devices to acquire the output data, is used as each of the recognition devices.
This invention relates to a recognition system configuration method for improving accuracy in multi-device recognition systems. The problem addressed is the variability and potential errors in recognition results when using multiple independent recognition devices, which can lead to inconsistent or unreliable outputs. The method involves configuring a recognition system where each recognition device is an ensemble recognition device. An ensemble recognition device combines recognition results from multiple individual recognition devices to produce a unified output. By using ensemble recognition devices as the primary recognition devices within the system, the method enhances the robustness and accuracy of the overall recognition process. The ensemble approach mitigates individual device errors by aggregating and refining results from multiple sources, ensuring more reliable output data. The configuration ensures that each recognition device in the system is capable of processing input data through multiple recognition pathways, cross-verifying results, and generating a consolidated output. This hierarchical ensemble structure improves the system's ability to handle diverse input conditions and reduces the likelihood of misrecognition. The method is particularly useful in applications requiring high accuracy, such as biometric identification, document processing, or automated quality control.
13. A recognition system configuration method of configuring the recognition system according to claim 4 , the recognition system configuration method comprising: causing the neural network for the first recognition task to learn with, as a learning data set, input data and output data of a learned recognition device that performs the first recognition task.
A recognition system configuration method involves training a neural network for a first recognition task using input and output data from a pre-trained recognition device that already performs the same task. The method leverages the learned behavior of an existing recognition system to configure a new neural network, ensuring it can replicate or improve upon the performance of the prior system. This approach reduces the need for extensive new training data by utilizing the output of a proven recognition device as a reference dataset. The neural network is trained to match or enhance the accuracy and efficiency of the pre-trained system, making it suitable for deployment in applications where the first recognition task is required. The method is particularly useful in scenarios where transferring knowledge from an established recognition system to a new neural network is desirable, such as in updating or scaling recognition technologies without starting from scratch. By using the learned outputs of the existing system, the method ensures consistency and reliability in the new neural network's performance.
14. The recognition system configuration method according to claim 13 , wherein an ensemble recognition device that unifies recognition results of a plurality of recognition devices to acquire the output data, is used as the recognition device.
Biometric or pattern recognition systems. This invention addresses the challenge of improving the accuracy and robustness of recognition systems. The method involves configuring a recognition system that utilizes an ensemble recognition device. This ensemble device is central to the configuration. The ensemble recognition device's function is to unify the recognition results obtained from multiple individual recognition devices. By combining the outputs of these diverse recognition devices, the ensemble device generates the final output data of the overall recognition system. This unification process aims to achieve a more reliable and accurate recognition outcome than could be obtained from any single recognition device alone. The configuration method thus focuses on setting up or defining a system where multiple recognizers contribute to a single, consolidated output.
15. A recognition system configuration method of configuring the recognition system according to claim 4 , the recognition system configuration method comprising: changing a structure of the generic neural network to cause a relationship between the input into the generic neural network and the output from the neural network for the first recognition task, to further approximate to a relationship between an input and an output of a learned recognition device that performs the first recognition task, and to cause a relationship between the input into the generic neural network and an output from a neural network for the second recognition task, to further approximate to a relationship between an input and an output of a learned recognition device that performs the second recognition task.
This invention relates to a method for configuring a recognition system, specifically a neural network-based system designed to perform multiple recognition tasks. The problem addressed is the need to adapt a generic neural network to closely mimic the input-output relationships of specialized, pre-trained recognition devices for different tasks. The method involves modifying the structure of the generic neural network to improve its performance on at least two distinct recognition tasks. For the first task, the network's structure is adjusted so that the relationship between its input and output more closely matches that of a learned recognition device trained specifically for that task. Similarly, the network is further modified to align its input-output relationship for the second task with that of another learned recognition device specialized in the second task. This approach allows a single neural network to approximate the behavior of multiple task-specific recognition systems, improving efficiency and adaptability in recognition applications. The method ensures that the generic network can handle diverse recognition tasks without requiring separate, task-specific networks.
16. A recognition system configuration method of configuring the recognition system according to claim 4 , the recognition system configuration method comprising: changing a structure of the neural network for the first recognition task, to cause a relationship between the input into the generic neural network and the output from the neural network for the first recognition task, to further approximate to a relationship between an input and an output of a learned recognition device that performs the first recognition task.
A recognition system configuration method involves adapting a neural network for a specific recognition task by modifying its structure to improve its performance. The method focuses on adjusting the neural network to better approximate the input-output relationship of a pre-trained recognition device that already performs the target task effectively. This approach ensures that the neural network can achieve higher accuracy or efficiency when applied to the same task. The configuration process may include altering layers, connections, or parameters within the neural network to refine its behavior. By fine-tuning the network's architecture, the method enables the system to closely mimic the performance of the learned recognition device, thereby enhancing its reliability and accuracy in executing the first recognition task. This technique is particularly useful in scenarios where pre-trained models or devices exist, allowing the neural network to leverage their learned relationships for improved results. The method ensures that the neural network's structure is optimized to handle the specific recognition task while maintaining compatibility with the broader recognition system.
Unknown
September 8, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.